With the construction of urban smart parking lots and the popularization of automatic toll collection systems at high-speed intersections, license plate recognition technology based on deep learning has been widely used. In order to solve a large number of blurred license plate character recognition in reality, a character free segmentation license plate character recognition algorithm based on improved CRNN+CTC(Recurrent Neural Network/Convolutional Neural Network+Connectionist Temporal Classification) is proposed. Firstly, the standard CNN in CRNN is changed into a micro-modified model of deeply separable convolutional network. Bi-directional long-term and short-term memory network is adopted in RNN, and CTC loss is introduced to train it. Secondly, in order to avoid the overfitting phenomenon in the training process, L2 regular term is added into the loss function and the training dataset is added. Finally, a batch normalization algorithm is introduced to accelerate the learning speed in the training process. Experimental results show that the proposed algorithm is applied to three experimental test sets. Experimental results show that compared with other methods based on complex environment, the proposed algorithm improves the average license plate recognition accuracy, recognition accuracy and speed on the three experimental test sets, and the robustness and generalization ability of the network are also stronger.